A dynamic trust assessment method for Internet of Vehicles

Abstract

A dynamic trust assessment method for an Internet of Vehicles (IoV) includes: acquiring a vehicle's sensor data, communication behavior data, and GPS data; then performing data cleaning, data synchronization, and feature extraction to obtain basic physical features and temporal dynamic features, with speed, communication frequency, speed variance, and communication burstiness used as the first, second, third, and fourth features, respectively; based on the four features, a roadside unit (ROU) calculates the vehicle's trust value using a weighted combination method, dynamically adjusting the weights of the first, second, third, and fourth features using a reinforcement learning network; the RSU dynamically adjusts the classification threshold based on the proportion of malicious nodes and the false alarm rate; if the vehicle's trust value is below the classification threshold, the vehicle is determined to be malicious; if the vehicle's trust value is not below the classification threshold, the vehicle is determined to be non-malicious. This invention enables more comprehensive, accurate, and efficient vehicle trust assessment.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle networking, and in particular to a dynamic trust assessment method for vehicle networking. Background Art

[0002] With the development of Internet of Vehicles (IoV) technology, data interaction between vehicles has become increasingly frequent. The trust value assessment of vehicles is crucial in ensuring the security and reliability of IoV systems.

[0003] However, existing vehicle trust assessment methods in the Internet of Vehicles have problems such as fixed feature weights, single feature utilization, rigid threshold setting, and lack of feature interaction, making it impossible to conduct comprehensive, accurate, and efficient vehicle trust assessment. Summary of the Invention

[0004] In view of this, the present invention provides a dynamic trust assessment method for an Internet of Vehicles to perform vehicle trust assessment in a more comprehensive, accurate and efficient manner.

[0005] A dynamic trust assessment method for an Internet of Vehicles, comprising:

[0006] Step S1: Acquire sensor data, communication behavior data, and GPS data of the vehicle, perform data cleaning, data synchronization, and feature extraction on the sensor data, communication behavior data, and GPS data to obtain basic physical characteristics and time series dynamic characteristics. The basic physical characteristics include speed and communication frequency, and the time series dynamic characteristics include speed variance and communication burstiness. Speed, communication frequency, speed variance, and communication burstiness are used as the first feature, second feature, third feature, and fourth feature, respectively.

[0007] Step S2: The roadside unit calculates the trust value of the vehicle based on the first feature, the second feature, the third feature, and the fourth feature using a weighted combination method, wherein the weights of the first feature, the second feature, the third feature, and the fourth feature are dynamically adjusted using a reinforcement learning network;

[0008] Step S3: The roadside unit dynamically adjusts the classification threshold based on the proportion of malicious nodes and the false alarm rate;

[0009] In step S4, if the trust value of the vehicle is lower than the classification threshold, the roadside unit determines that the vehicle is a malicious vehicle; if the trust value of the vehicle is not lower than the classification threshold, the roadside unit determines that the vehicle is a non-malicious vehicle.

[0010] The dynamic trust assessment method for the Internet of Vehicles (IoV) provided by the present invention integrates sensor data, communication behavior data, and GPS data, addressing the problem of missing feature interactions. The method uses speed and communication frequency as fundamental physical features, speed variance and communication burstiness as temporal dynamic features, and calculates the vehicle's trust value using a weighted combination approach, addressing the problem of single feature utilization. The method uses a reinforcement learning network to dynamically adjust the weights of speed, communication frequency, speed variance, and communication burstiness, and dynamically adjusts the classification threshold based on the proportion of malicious nodes and the false alarm rate. This creates a dual dynamic adjustment mechanism of dynamic weight allocation and adaptive classification threshold adjustment, addressing the issues of fixed feature weights and rigid threshold settings. It enables high-precision malicious node detection, resulting in more comprehensive, accurate, and efficient vehicle trust assessment. Experiments show that the method can improve the F1 value of malicious node detection to 94.22% and the accuracy to 97.74% in real IoV scenarios, significantly outperforming existing solutions. This method is applicable to scenarios such as V2X communication and platoon collaborative driving, and can significantly improve the security and reliability of IoV systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 A flowchart of a method for dynamic trust assessment of an Internet of Vehicles provided by an embodiment of the present invention;

[0012] Figure 2 This is a graph showing the dynamic adjustment of weights in the present invention;

[0013] Figure 3 A graph showing the dynamic adjustment of the classification threshold value according to the present invention;

[0014] Figure 4 This is a graph showing the changes in the F1 value of the present invention. DETAILED DESCRIPTION

[0015] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the embodiments of the present invention, and should not be construed as limiting the present invention.

[0016] See also Figure 1 , an embodiment of the present invention provides a method for dynamic trust assessment of an Internet of Vehicles, comprising steps S1 to S4:

[0017] Step S1: Acquire the vehicle's sensor data, communication behavior data, and GPS data, and perform data cleaning, data synchronization, and feature extraction on the sensor data, communication behavior data, and GPS data to obtain basic physical characteristics and time-series dynamic characteristics. The basic physical characteristics include speed and communication frequency, and the time-series dynamic characteristics include speed variance and communication burstiness. Speed, communication frequency, speed variance, and communication burstiness are used as the first feature, second feature, third feature, and fourth feature, respectively.

[0018] Among them, vehicle-mounted sensors and communication modules (such as CAN bus, OBD-II interface) are used to collect sensor data, GPS location information and communication behavior data.

[0019] Sensor data includes the vehicle's speed, acceleration, and steering angle; communication behavior data includes the time interval between vehicle messages, the logical consistency of messages, the source and target nodes of messages; GPS data includes real-time coordinates, timestamps, and driving routes.

[0020] When cleaning data, use Python libraries (such as NumPy) for batch processing to remove noise data and missing values and improve data quality.

[0021] Specifically, for speed, acceleration, and steering angle, the rationality and continuity of the data must be ensured. The upper and lower limits of this type of data are easy to determine, and it is easy to identify obvious anomalies in the physical quantity. Therefore, upper and lower limit constraints are used to eliminate outliers, and interpolation is used to supplement missing values or eliminated outliers. Then, a sliding average is used to process speed, and a median filter is used to process acceleration and steering angle.

[0022] Among them, when the sliding average processing speed is adopted, the following formula is satisfied:

[0023] ;

[0024] in, is the smoothed data, Indicates the total size of the window, , is the sampling point index, Indicates the length of one side of the sliding window, Indicates the The original data.

[0025] Median filtering is to take the median value in the sliding window to replace the value of the window center point.

[0026] During data cleaning, for GPS data, GPS data cleaning must ensure the continuity of GPS data and eliminate jumps and drifts. This embodiment uses the upper and lower limit constraint method to eliminate outliers, uses the trajectory interpolation algorithm to correct GPS abnormal deviation points, and uses speed and distance thresholds to detect jumps;

[0027] During data cleaning, communication behavior data is classified and cannot be corrected through simple numerical interpolation. Therefore, the denoising method for communication behavior data is as follows: first, rule filtering is performed to define a set of legal values and remove values that are not in the legal value set; second, based on the reasonable logic of the behavior sequence, behaviors that do not meet expectations are detected and marked; finally, outliers are corrected, either by replacing them with the most recent valid behavior or filling them with "unknown".

[0028] When synchronizing data, the vehicle uses Python libraries (such as NumPy) for batch processing and performs timestamp alignment and frame alignment.

[0029] The process of timestamp alignment is: unify the clock source (such as the vehicle clock), unify the timestamps of different data sources (such as sensors, GPS, and communication modules) to a standard time base, and use linear interpolation to upsample low-frequency data to high-frequency data for data with different sampling frequencies (such as 1Hz for GPS and 100Hz for sensors). Because there is a lot of high-frequency data and less low-frequency data, upsampling low-frequency data to high-frequency data is essentially to fill in some data. Specifically: determine the interval between low-frequency data points; for high-frequency time points , find the corresponding two low-frequency time points 、 , use linear interpolation to calculate the high frequency value, the calculation formula is as follows:

[0030]

[0031] in, represents high frequency values, and are two adjacent points in the low-frequency data.

[0032] The frame alignment process aggregates speed, acceleration, steering angle, GPS data, and communication behavior data within the same time window into a complete vector based on aligned timestamps. After data cleaning and synchronization, the vehicle sends the complete vector to the roadside unit (RSU), which performs subsequent feature extraction.

[0033] During feature extraction, the vehicle sends the complete vector to the roadside unit, which uses a feature engineering toolkit (such as scikit-learn) to extract basic physical features and time-series dynamic features. The basic physical features include speed and communication frequency, and the time-series dynamic features include speed variance and communication burstiness. Speed, communication frequency, speed variance, and communication burstiness are used as the first, second, third, and fourth features, respectively.

[0034] The speed variance is used to identify abnormal acceleration and deceleration behaviors (such as random speed disturbances of malicious nodes). The speed variance is calculated as follows:

[0035] ;

[0036] in, express The velocity variance at each moment, Indicates the sliding window size for speed behavior analysis, express The speed of time, Indicates average speed.

[0037] Communication burstiness is used to detect sudden changes in communication frequency. The calculation formula for communication burstiness is:

[0038] ;

[0039] in, express Communication burstiness index at the moment, Indicates the sliding window size for communication behavior analysis, express The communication frequency at the moment, express The communication frequency at the moment.

[0040] Finally, each feature needs to be standardized so that its value range is limited to [0,1], and we get Multi-scale feature vector at the moment , , 、 、 、 Represents vehicles exist The first, second, third, and fourth characteristics of a moment.

[0041] Step S2, based on the first feature, the second feature, the third feature and the fourth feature, the trust value of the vehicle is calculated using a weighted combination method, wherein the weights of the first feature, the second feature, the third feature and the fourth feature are dynamically adjusted using a reinforcement learning network.

[0042] Among them, in the process of calculating the trust value of the vehicle using the weight combination method, the following formula is satisfied:

[0043] ;

[0044] in, Indicates vehicle exist The trust value of the moment, Indicates vehicle exist The moment feature, , express Moment The weight of the feature, It is the smoothness control value, which is generally taken as . , 、 、 、 They are The weights of the first, second, third, and fourth features at each moment.

[0045] Among them, the roadside unit dynamically adjusts the classification threshold through the reinforcement learning network.

[0046] The reinforcement learning network adopts a dual-depth Q-network architecture, which includes an online Q-network and a target Q-network, and improves stability through experience replay and delayed updates.

[0047] The process of dynamically adjusting the weights of the first feature, the second feature, the third feature, and the fourth feature using the reinforcement learning network specifically includes:

[0048] Step S21, state space construction: state vector ,in, for The multi-scale feature vector at time , , 、 、 、 Represents vehicles exist The first, second, third, and fourth characteristics of a moment, for The weight vector at the moment, , 、 、 、 Respectively The weights of the first, second, third, and fourth features at each moment, yes The classification threshold at time t, It is a historical window statistic (such as the proportion of malicious nodes in the past 10 seconds and the average confidence level);

[0049] Step S22, action space design: Each feature corresponds to an independent action space, allowing fine-grained adjustment of weights. , 、 、 、 Represent the adjustment amount of the weight of the first feature, the second feature, the third feature, and the fourth feature respectively, ;

[0050] Step S23: reward function design. Based on the trust error and malicious node detection performance, the reward function is defined as:

[0051] ;

[0052] ;

[0053]

[0054] in, is the reward function, is the mean square error between the predicted trust value and the true trust value, is the performance reward detection value, and is the weight coefficient, It is The true trust value of the sample, It is The predicted confidence value of the sample, is the total number of samples, is the accuracy rate, which indicates the proportion of malicious nodes detected by RSU that are actually malicious nodes. , Indicates the number of malicious nodes that are correctly detected, Indicates the number of normal nodes that are mistakenly detected as malicious nodes, is the recall rate, which represents the proportion of correctly detected malicious nodes among all actual malicious nodes. , Indicates the number of malicious nodes that are mistakenly detected as normal nodes;

[0055] Step S24, network update rules:

[0056] The online Q network is updated every first preset time (100 milliseconds in this embodiment) and uses gradient descent to optimize the loss function. The loss function for:

[0057] ;

[0058]

[0059] in, represents an empirical data point for reinforcement learning, is the current state, Is in state The following actions are taken, Is to perform an action After receiving the reward, is the next state reached after executing the action, Indicates the sampling of the experience replay buffer Perform expectation calculation, that is, calculate the mean squared error (MSE) loss on multiple samples to obtain a more stable training signal. represents the Q value predicted by the current online Q network, represents the target Q value, is a discount factor that controls the impact of future rewards, Indicates using the target Q network to calculate the next state The Q value of the optimal action.

[0060] Gradient descent is used to optimize the parameters to minimize the error between the predicted Q value and the target Q value.

[0061] The target Q network synchronizes the parameters of the online Q network once every second preset time (1000 milliseconds in this embodiment). The target network is represented as:

[0062] ;

[0063] in, are the parameters of the target Q network, are the parameters of the online Q network, is the update speed control value (0.001 in this embodiment).

[0064] In this embodiment, after the weights of the first feature, the second feature, the third feature, and the fourth feature are dynamically adjusted using a reinforcement learning network, the weights of the first feature, the second feature, the third feature, and the fourth feature meet the non-negativity and normalization conditions, and the expression is:

[0065] ;

[0066] in, express Moment The weight of the feature, is the standard weight adjustment value, is a constant term (in this embodiment, 10 -8 ), Indicates taking the maximum value.

[0067] In this embodiment, the dynamic adjustment process of weights can be found in Figure 2 ,from Figure 2It can be seen that as the training is continuously adjusted, the weights gradually tend to be stable, indicating that the present invention can effectively perform dynamic adjustment of feature weights.

[0068] In step S3, the roadside unit dynamically adjusts the classification threshold based on the proportion of malicious nodes and the false alarm rate.

[0069] Among them, based on the dual feedback mechanism of malicious node ratio and false alarm rate, the classification threshold is dynamically adjusted, and the expression is:

[0070] ;

[0071] ;

[0072] in, express The classification threshold at time t, is the truncation function, express The classification threshold at time t, for The adjustment amount of the moment classification threshold, is the influence coefficient of the proportion of malicious nodes, is the benchmark for the proportion of malicious nodes, is the influence coefficient of false alarm rate, is the proportion of malicious nodes detected, , Indicates the number of nodes currently detected as malicious. Indicates the total number of nodes currently participating in the detection. for The false alarm rate at each moment, , Indicates the number of normal nodes that are correctly detected. function, so Always between 0.35 and 0.65 to prevent over-adjustment.

[0073] In this embodiment, the dynamic adjustment process of the classification threshold can be referred to Figure 3 ,from Figure 3 It can be seen that the classification threshold fluctuates slowly between multiple values and finally converges to a stable value (around 0.43), indicating that the present invention can automatically find the optimal classification boundary according to the data feature distribution.

[0074] Step S4: If the trust value of the vehicle is lower than the classification threshold, the vehicle is determined to be a malicious vehicle; if the trust value of the vehicle is not lower than the classification threshold, the vehicle is determined to be a non-malicious vehicle.

[0075] Among them, RSU judges vehicles with trust values lower than the classification threshold as malicious vehicles and will not communicate with the malicious vehicles subsequently.

[0076] In addition, this embodiment supports lightweight online updates with an update cycle of ≤50ms to ensure adaptability to dynamic environments.

[0077] The present invention was tested below.

[0078] To verify the effectiveness of the proposed method in the Internet of Vehicles, a simulated data set was constructed and a complete comparative experiment was conducted. The specific settings are as follows:

[0079] 1. Experimental Background

[0080] This experiment simulates the normal and malicious behaviors of multiple vehicle nodes in the Internet of Vehicles environment, generates time series feature samples containing multiple physical characteristics and communication behaviors, and evaluates the recognition ability and robustness of the method of the present invention for malicious nodes in a dynamic environment.

[0081] 2. Data Generation

[0082] Total number of samples: 5000 time window samples.

[0083] Malicious node ratio: 20% (i.e., 1,000 malicious samples).

[0084] Feature dimensions: 4 dimensions in total, namely:

[0085] speed: vehicle speed, normal distribution N(60, 15), clipped to the range [0, 120];

[0086] acceleration: vehicle acceleration, normal distribution N(0, 2);

[0087] comm_freq: communication frequency, following Poisson distribution λ=12;

[0088] gps_deviation: GPS deviation, which follows the normal distribution N(0, 0.1) and takes the absolute value.

[0089] Malicious nodes simulate disturbances:

[0090] Vehicle speed deviation (±25 km / h), manifested as sudden acceleration or deceleration;

[0091] The communication frequency is set to Poisson distribution (λ = 25) to simulate frequent communication behavior;

[0092] The GPS offset is additionally perturbed with a standard deviation of N(0, 0.3) to simulate location spoofing;

[0093] Standardization: Use StandardScaler to standardize all features to improve the stability of model training.

[0094] Evaluation method: 5-fold time series cross-validation, compared with a fixed weight method (weighted equally, threshold of 0.5).

[0095] 3. Method design

[0096] (1) The method of the present invention:

[0097] Weights are updated using the momentum optimizer (momentum factor 0.9) and a learning rate of 0.1.

[0098] Calculate the optimal classification threshold using the method of the present invention;

[0099] Threshold search range: [0.35, 0.65];

[0100] Weight normalization is performed to ensure non-negativity and a sum of 1;

[0101] The maximum number of training rounds is 200 rounds. If the F1 score does not improve after 20 rounds, it will be terminated early.

[0102] (2) Comparison method: using the baseline method, namely the fixed weight threshold method:

[0103] All features are equally weighted (1 / 4);

[0104] The classification threshold is fixed at 0.5;

[0105] There is no dynamic adjustment process.

[0106] 4. Training and Validation

[0107] Data segmentation:

[0108] Use TimeSeriesSplit to perform 5-fold time series cross-validation to ensure that the training set is earlier than the validation set, which conforms to the temporal logic.

[0109] Each round of training performs weight learning and threshold tuning on the currently divided training set and evaluates it on the validation set.

[0110] Evaluation Metrics:

[0111] Accuracy

[0112] Recall

[0113] F1 Score.

[0114] Statistical tests:

[0115] The paired T test was used to verify the significance of the F1 values of the method of the present invention and the comparative method.

[0116] 5. Experimental results

[0117] The variation of the F1 value of the present invention can be found in Figure 4 ,from Figure 4 It can be seen that the F1 score of the present invention increases steadily and eventually tends to be stable or fluctuates slightly, indicating that the training process of the present invention is effective and the performance is gradually improved; there is no overfitting (no significant drop in F1 in the late stage of training is observed); the training process converges well, and the reinforcement learning strategy is stable and effective.

[0118] The comparison between the method of the present invention and the comparative method is shown in Table 1:

[0119] Table 1

[0120]

[0121] The T-test result of this experiment is 0.0000. As can be seen from Table 1, the performance of the present invention is significantly better than the baseline method.

[0122] In summary, the dynamic trust assessment method for the Internet of Vehicles (IoV) according to the above-mentioned embodiment integrates sensor data, communication behavior data, and GPS data, solving the problem of missing feature interactions. The present invention uses speed and communication frequency as basic physical features, speed variance and communication burstiness as temporal dynamic features, and calculates the trust value of the vehicle using a weighted combination method, solving the problem of single feature utilization. The present invention uses a reinforcement learning network to dynamically adjust the weights of speed, communication frequency, speed variance, and communication burstiness, and dynamically adjusts the classification threshold based on the proportion of malicious nodes and the false alarm rate. This forms a dual dynamic adjustment mechanism of dynamic weight allocation and adaptive classification threshold adjustment, solving the problems of fixed feature weights and rigid threshold settings. It can achieve high-precision malicious node detection, thereby performing vehicle trust assessment more comprehensively, accurately, and efficiently. Experiments show that the present invention can improve the F1 value of malicious node detection to 94.22% and the accuracy to 97.74% in real IoV scenarios, significantly outperforming existing technical solutions. This method is suitable for scenarios such as V2X communication and convoy cooperative driving, and can significantly improve the security and reliability of IoV systems.

[0123] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A dynamic trust assessment method for an Internet of Vehicles, characterized in that: include: Step S1: Acquire the vehicle's speed, acceleration, steering angle, communication behavior data, and GPS data. The communication behavior data includes the time interval for the vehicle to send messages, the logical consistency of the sent messages, and the source and target nodes of the messages. Perform data cleaning, data synchronization, and feature extraction on the speed, acceleration, steering angle, communication behavior data, and GPS data to obtain basic physical characteristics and time series dynamic characteristics. The basic physical characteristics include speed and communication frequency. The time series dynamic characteristics include speed variance and communication burstiness. Communication burstiness is used to detect sudden changes in communication frequency. Speed, communication frequency, speed variance, and communication burstiness are used as the first feature, second feature, third feature, and fourth feature, respectively. Step S2: The roadside unit calculates the trust value of the vehicle based on the first feature, the second feature, the third feature, and the fourth feature using a weighted combination method, wherein the weights of the first feature, the second feature, the third feature, and the fourth feature are dynamically adjusted using a reinforcement learning network; Step S3: The roadside unit dynamically adjusts the classification threshold based on the proportion of malicious nodes and the false alarm rate; Step S4: If the trust value of the vehicle is lower than the classification threshold, the roadside unit determines that the vehicle is a malicious vehicle; if the trust value of the vehicle is not lower than the classification threshold, the roadside unit determines that the vehicle is a non-malicious vehicle; In step S2, when calculating the trust value of the vehicle using the weighted combination method, the following formula is satisfied: ； in, Indicates vehicle exist The trust value of the moment, Indicates vehicle exist The moment feature, express Moment The weight of the feature, is the smoothness control value, , 、 、 、 They are The weights of the first, second, third, and fourth features at each moment.

2. The method for dynamic trust assessment of Internet of Vehicles according to claim 1, characterized in that: In step S1, the GPS data includes real-time coordinates, timestamp, and driving route.

3. The method for dynamic trust assessment of Internet of Vehicles according to claim 2, characterized in that: In step S1, during data cleaning, for speed, acceleration, and steering angle, the upper and lower limit constraints are used to eliminate outliers, and the interpolation method is used to supplement missing values or eliminated outliers. Then, the speed is processed by sliding average, and the acceleration and steering angle are processed by median filtering. Among them, when the sliding average processing speed is adopted, the following formula is satisfied: ； in, is the smoothed data, Indicates the total size of the window, , is the sampling point index, Indicates the length of one side of the sliding window, Indicates the Original data; During data cleaning, for GPS data, the upper and lower limit constraints are used to eliminate outliers, the trajectory interpolation algorithm is used to correct GPS abnormal deviation points, and the speed and distance thresholds are used to detect jump points; During data cleaning, for communication behavior data, rule filtering is first performed to define a set of legal values and remove values that are not in the set. Secondly, based on the reasonable logic of the behavior sequence, behaviors that do not meet expectations are detected and marked. Finally, outliers are corrected.

4. The method for dynamic trust assessment of Internet of Vehicles according to claim 3, characterized in that: In step S1, when data is synchronized, timestamp alignment and frame alignment are performed respectively; The timestamp alignment process is as follows: unify the clock source, unify the timestamps of different data sources to a standard time base, and use linear interpolation to upsample low-frequency data to high-frequency data for data with different sampling frequencies. The linear interpolation method satisfies the following equation: in, represents high frequency values, and are two adjacent points in the low-frequency data, is a high-frequency time point; The frame alignment process is to aggregate the velocity, acceleration, steering angle, GPS data, and communication behavior data into a set of complete vectors within the same time window based on the aligned timestamps.

5. The method for dynamic trust assessment of Internet of Vehicles according to claim 4, characterized in that: In step S1, during feature extraction, the vehicle sends a complete vector to the roadside unit, which uses a feature engineering toolkit to extract basic physical features and time-series dynamic features. The basic physical features include speed and communication frequency, and the time-series dynamic features include speed variance and communication burstiness. Speed, communication frequency, speed variance, and communication burstiness are used as the first feature, second feature, third feature, and fourth feature, respectively. The calculation formula of velocity variance is: ； in, express The velocity variance at each moment, Indicates the sliding window size for speed behavior analysis, express The speed of time, Indicates average speed; The calculation formula for communication burstiness is: ； in, express Communication burstiness index at the moment, Indicates the sliding window size for communication behavior analysis, express The communication frequency at the moment, express The communication frequency at the moment.

6. The method for dynamic trust assessment of Internet of Vehicles according to claim 5, characterized in that: In step S2, the reinforcement learning network adopts a dual-depth Q-network architecture, including an online Q-network and a target Q-network, and improves stability through experience replay and delayed update; The process of dynamically adjusting the weights of the first feature, the second feature, the third feature, and the fourth feature using the reinforcement learning network specifically includes: Step S21, state space construction: state vector ,in, for The multi-scale feature vector at time , , 、 、 、 Represents vehicles exist The first, second, third, and fourth characteristics of a moment, for The weight vector at the moment, , 、 、 、 Respectively The weights of the first, second, third, and fourth features at each moment, yes The classification threshold at time t, is a historical window statistic; Step S22, action space design: each feature corresponds to an independent action space, and the weight adjustment action , 、 、 、 Respectively represent the adjustment amounts of the weights of the first feature, the second feature, the third feature, and the fourth feature; Step S23: reward function design. Based on the trust error and malicious node detection performance, the reward function is defined as: ； ； in, is the reward function, is the mean square error between the predicted trust value and the true trust value, is the performance reward detection value, and is the weight coefficient, It is The true trust value of the sample, It is The predicted confidence value of the sample, is the total number of samples, is the accuracy, , Indicates the number of malicious nodes that are correctly detected, Indicates the number of normal nodes that are mistakenly detected as malicious nodes, is the recall rate, , Indicates the number of malicious nodes that are mistakenly detected as normal nodes; Step S24, network update rules: The online Q network is updated every first preset time and uses gradient descent to optimize the loss function. The loss function for: ； in, represents an empirical data point for reinforcement learning, is the current state, Is in state The following actions are taken, Is to perform an action After receiving the reward, is the next state reached after executing the action, Indicates the sampling of the experience replay buffer Perform expectation calculations, represents the Q value predicted by the current online Q network, represents the target Q value, is the discount factor, Indicates using the target Q network to calculate the next state The Q value of the optimal action; The target Q network synchronizes the parameters of the online Q network every second preset time. The target network is expressed as: ； in, are the parameters of the target Q network, are the parameters of the online Q network, It is the update speed control value.

7. The method for dynamic trust assessment of Internet of Vehicles according to claim 6, characterized in that: In step S2, after the weights of the first feature, the second feature, the third feature, and the fourth feature are dynamically adjusted using the reinforcement learning network, the weights of the first feature, the second feature, the third feature, and the fourth feature meet the non-negativity and normalization conditions, and the expression is: ； in, express Moment The weight of the feature, is the standard weight adjustment value, is a constant term, Indicates taking the maximum value.

8. The method for dynamic trust assessment of Internet of Vehicles according to claim 7, characterized in that: Step S3 satisfies the following formula: ； ； in, express The classification threshold at time t, is the truncation function, express The classification threshold at time t, for The adjustment amount of the moment classification threshold, is the influence coefficient of the proportion of malicious nodes, is the benchmark for the proportion of malicious nodes, is the influence coefficient of false alarm rate, is the proportion of malicious nodes detected, , Indicates the number of nodes currently detected as malicious. Indicates the total number of nodes currently participating in the detection. for The false alarm rate at each moment, , Indicates the number of normal nodes that are correctly detected.

Images